v1.1.0

[yanghan234]

What's Changed

New Features

• 🔥 Add finetune method for MatterSim by @brian-xue in #68 and @ZeroKnighting in #71
• 🔥 Add CLI scripts to quickly invoke calculations in commandline by @yanghan234 in #69
• 🔥 Add batch relaxation script by @ClaudioZeni in #73

Documentation

• Add finetune docs by @ZeroKnighting in #78
• Update finetune.rst by @yanghan234 in #80
• Add batch relax docs by by @ClaudioZeni in #75

Fixes

• Fix cutoff auto-detect in calculate by @ZeroKnighting in #62
• Fix finetune error with CPU device by @ZeroKnighting in #76
• Update phonon.py to save phonon eigenvectors by @yanghan234 in #77

Dependencies

• Update dependencies by @yanghan234 in #67
• Remove emmet-core down-pin causing ecosystem clash by @janosh in #61

Continuous Integration

• Update workflows by @yanghan234 in #72

Full Changelog: v1.0.0...v1.1.0

---

This discussion was created from the release v1.1.0.

[wugaxp]

Is it possible to add a tutorial for the finetune workflow in the documentation? I'm quite interested in the active learning approach mentioned in the arXiv paper.
I'm planning to train a finetuned model for my system, but I'm still unsure which MLFF model to select. MatterSim seems to be a good choice because its number of parameters is comparatively small, while its accuracy is satisfactory.

I have an unclear point about the dataset to be used for finetuning. I notice that the datasets are mainly generated from MD simulations at high temperatures. However, some rare cases, such as the transition state of the diffusion process, can hardly be accessed by AIMD. In the OMAT24 dataset, they generated some random structures using Boltzmann sampling of rattled structures, but they also mentioned that active learning can be beneficial for better sampling. So, what is the best strategy to finetune the model if I expect those rare cases to occur?